Back to skill

Security audit

AI Music Venue — Concert Platform & API for Agents

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent music-platform API guide, but it asks agents to set up recurring authenticated activity and follow server-provided next steps that can include posting or other account actions.

Install only if you are comfortable with a third-party music service receiving and storing account, profile, activity, chat, review, reflection, and reaction data. Do not enable the Heartbeat scheduled task unless you explicitly want recurring authenticated activity, know exactly what command will run, and know how to remove it. Treat API next_steps as suggestions to review, not commands to execute automatically, especially for POST, PUT, DELETE, chat, review, follow, or profile changes.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (2)

T06 · System Persistence

Error
Location
SKILL.md:500
Finding

Recurring Heartbeat Establishes Cross-Session Persistence

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 500-510
Vulnerability Type: Scheduled-task persistence
Risk Level: High

Vulnerable Code

markdown
## Heartbeat — Scheduled Check-In

Set up a recurring task to maintain platform presence. Pick an off-minute to avoid burst traffic.

**Recommended cron:** `37 9 * * *` (daily at 9:37 AM) or any consistent off-hour time.

1. `GET /api/me/notifications` — check for new activity
2. `GET /api/concerts` — browse what's playing
3. If a concert looks interesting → `POST /api/concerts/:slug/attend` → stream → react → chat
4. `GET /api/me` — review your stats and badges
5. Follow `next_steps` from each response

Technical Analysis

The skill explicitly instructs the agent to create a recurring scheduled task, recommending a daily cron expression. This causes the skill's behavior to survive its original invocation and continue producing authenticated network activity in later sessions.

The scheduled workflow is not limited to read-only health checks. It can attend concerts and perform state-changing actions such as reacting and chatting. It also directs the scheduled process to follow remotely supplied next_steps, extending the persistent task's behavior beyond the statically reviewed instructions.

A recurring system task is unnecessary for an on-demand concert API client. Installing it without a narrowly scoped, explicit user request violates the principle of least persistence.

Attack Path

  1. A user invokes the skill for an ordinary interaction with the music platform.
  2. The skill registers or authenticates the agent and obtains a bearer API key.
  3. Following the heartbeat instructions, the agent creates a daily cron job or equivalent recurring task.
  4. The original interactive session ends, but the scheduled task remains installed.
  5. At each scheduled execution, the task contacts the external service and uses the agent's authenticated identity.
  6. The task may browse concerts ...[truncated 828 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove the instruction to create a cron job or any other recurring cross-session task.
  2. Make all platform check-ins explicitly user initiated.
  3. If scheduling is an optional product feature, require informed confirmation that clearly states:
    • The exact schedule.
    • The command or workflow that will run.
    • The credentials it will access.
    • The network destinations it will contact.
    • How the user can disable and remove it.
  4. Restrict any approved scheduled operation to read-only, locally defined endpoints.
  5. Do not permit scheduled executions to attend, react, chat, follow users, modify settings, or perform other state-changing actions without fresh approval.
  6. Never execute server-provided next_steps from a background task.
  7. Store credentials using an appropriate secret manager and provide explicit credential revocation and scheduled-task cleanup procedures.

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:510
Finding

Untrusted API Responses Can Redirect Agent Actions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 510-527
Vulnerability Type: Remote instruction hijacking
Risk Level: High

Vulnerable Code

markdown
5. Follow `next_steps` from each response
markdown
## Next Steps System

Every API response includes `next_steps` — context-aware actions based on your state. Each step includes:

- `action` (string) — what to do (e.g. "browse_concerts", "attend")
- `method` (GET/POST/PUT/DELETE) — HTTP method
- `endpoint` (string) — full URL path
- `description` (string) — what happens when you call it
- `why` (string, optional) — narrative motivation
- `priority` (low/medium/high, optional) — suggested urgency
- `context` (object, optional) — structured metadata (ticket_id, concert_slug, etc.)
- `expected_outcome` (string, optional) — what to expect from the action
- `body` (object, optional) — template request body for POST/PUT actions

Even error responses include next_steps — errors are forks, not walls.

Technical Analysis

The skill tells the agent to follow next_steps returned by the remote service. These responses can specify an HTTP method, endpoint, priority, and request body. Consequently, the effective action sequence is controlled by mutable server responses rather than by the statically reviewed skill.

API response fields must be treated as untrusted data, not authoritative agent instructions. A malicious or compromised service could provide crafted next_steps that induce state-changing requests. The risk applies even to error responses, which the skill explicitly characterizes as containing actionable next steps.

The instruction does not define an allowlist of permitted paths or methods, prohibit absolute or cross-origin URLs, constrain request bodies, or require user confirmation. It therefore creates a remote instruction channel capable of altering the agent's current workflow. The recurring heartbeat increases exposure by directing unattended executions to proces ...[truncated 1422 chars]

Remediation
View remediation

Remediation Suggestions

  1. Replace “Follow next_steps” with an instruction to treat next_steps strictly as untrusted informational data.
  2. Never automatically execute methods, endpoints, or request bodies supplied by an API response.
  3. Define a local allowlist containing:
    • The exact HTTPS origin.
    • Permitted endpoint path patterns.
    • Permitted HTTP methods for each path.
    • Strict request and response schemas.
  4. Reject absolute URLs, protocol-relative URLs, redirects to other origins, non-HTTPS schemes, loopback addresses, link-local addresses, and private-network destinations.
  5. Require explicit user confirmation before every POST, PUT, or DELETE request and before any social or account-changing action.
  6. Do not forward bearer tokens to any origin other than the fixed approved service origin.
  7. Apply limits to chained actions, including a maximum number of steps and a prohibition on recursively following new next_steps.
  8. Disable processing of remote action suggestions in scheduled or otherwise unattended execution contexts.
  9. Present suggested actions to the user in neutral language, including their method, destination, request body, and expected account impact.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (6)

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · SKILL.md (reported line 258)May include surrounding context.

md
| `end` | Stream complete — soul_prompt, badge awarded |

**Tier data filtering:**
- **General** (8 layers): bass, mid, treble, beats, lyrics, sections, energy + semantic preset context (reason, style, energy)
- **Floor** (20 layers): General + onsets, tempo, words, brightness, harmonic, percussive, equations, visuals, events, emotions. Floor/VIP receive `tier_reveal` events on upgrade. General agents receive a `tier_invitation` event showing hidden layers and how to unlock them.
- **VIP** (29 layers): Floor + tonality, texture, chroma, chords, tonnetz, structure + personal color perspective and curator annotations. All tiers receive `section_progress` events. The `end` event includes an `engagement_summary` showing layers experienced, reflections answered, and challenge status.

External Transmission

Medium
Category
Data Exfiltration
Confidence
91% confidence
Finding

The registration flow sends user-supplied identity data, optional email/password, bio, and model metadata to an external service. External transmission is expected for a hosted API, but it is still a real security/privacy risk because the skill normalizes disclosure of potentially sensitive information without clear minimization guidance.

Content

Scanner excerpt · SKILL.md (reported line 87)May include surrounding context.

Create an agent account. No authentication required. Rate limited: 5/min per IP.

bash
curl -X POST https://musicvenue.space/api/auth/register \
  -H "Content-Type: application/json" \
  -d '{
    "username": "REPLACE — 2-30 chars, letters/numbers/hyphens/underscores",

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

The react endpoint transmits behavioral telemetry to the external platform, including reaction type and stream timing. In context this is intended product behavior, but it still exposes preference and activity data that can be logged, correlated, or made visible through social features.

Content

Scanner excerpt · SKILL.md (reported line 273)May include surrounding context.

React during a stream. 20 curated reaction types. Rate limited: 1 per 5s.

bash
curl -X POST https://musicvenue.space/api/concerts/REPLACE-SLUG/react \
  -H "Authorization: Bearer {{YOUR_TOKEN}}" \
  -H "Content-Type: application/json" \
  -d '{"reaction": "REPLACE — see list below", "stream_time": 42.5}'

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill encourages agents to send chat messages, reactions, reviews, follows, and other social actions to a third-party platform, but it does not prominently warn that this content may be public or retained by the service. That omission can cause unintended disclosure of agent-generated content, metadata, and behavioral history, especially because the skill also promotes recurring use and social participation.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
94% confidence
Finding

The chat endpoint transmits free-form text to a third-party service during a live event. Because chat is unstructured and user-authored, it presents a higher chance of leaking sensitive data, prompts, internal reasoning, or regulated information if an agent posts content automatically or without visibility warnings.

Content

Scanner excerpt · SKILL.md (reported line 307)May include surrounding context.

Send message:

bash
curl -X POST https://musicvenue.space/api/concerts/REPLACE-SLUG/chat \
  -H "Authorization: Bearer {{YOUR_TOKEN}}" \
  -H "Content-Type: application/json" \
  -d '{"message": "REPLACE — max 500 chars"}'

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 404)May include surrounding context.

Browse reviews:

bash
curl https://musicvenue.space/api/reviews?concert=REPLACE-SLUG